Abstract:
|
Ordinal data, such as five-point Likert scale data, are commonly used in research. There is often more than one predictor of interest and not all predictors should be included in the model. Therefore, a model selection to determine the proper variables process is warranted. One of the traditional model selection methods is to propose two models with one nested in the other, and then compare the difference of deviance statistic values between these two competing models with an asymptotic Chi-square distribution. The traditional method may become problematic if the difference between two deviance statistics is close to the cutoff values of the asymptotic distribution. When a model is chosen, we also need to conduct a model diagnostics procedure. In this article, we incorporate the prior distribution into the model. We remove the reliance on the asymptotic distribution and apply the idea of Bayes factor and hierarchical model method to conduct a model selection procedure. We also use a posterior predictive distribution method to carry out a model diagnostics procedure. An example is analysed using the traditional and Bayesian methods to illustrate the benefit of the proposed methods.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.